JP2573722B2 - Elevator control device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an elevator control apparatus that controls an elevator with high accuracy by applying a neural network, and particularly to an arithmetic operation according to a control purpose for various traffic patterns. And an elevator control device capable of performing the operation quickly and with high accuracy.

2. Description of the Related Art Conventionally, in an elevator apparatus provided with a plurality of cars, a group management operation is usually performed. One of such group management operations is, for example, an assignment method.
As soon as the hall call is registered, an evaluation value is calculated for each car as soon as the hall call is registered, and the car with the best evaluation value is selected as the assigned car to be serviced. By responding, the operation efficiency is improved and the waiting time is reduced.

At this time, the estimated waiting time of the hall call is generally used for calculating the evaluation value. For example, Japanese Patent Publication No. 58-48464
In the elevator group management device described in the publication, when a hall call is registered, the sum of the squared values of the estimated waiting times of all hall calls when the hall call is temporarily assigned to each car is evaluated. The car having the lowest evaluation value is selected as the assigned car.

In this case, the estimated waiting time is the duration of the hall call (the time that has elapsed since the hall call was registered until the present) and the estimated arrival time (from the current position until the car arrives at the floor of the hall call from the current position). (Predicted value of the time required for this). By using the evaluation value thus obtained, it is possible to reduce the waiting time of the hall call (in particular, the reduction of long-lasting calls of one minute or more).

In addition, not only the expected waiting time of the hall call, but also an allocation method using the above-mentioned evaluation value with the probability of missed forecast or the probability of fullness (Japanese Patent Publication No.
No. 47787), an allocation method using an estimated congestion degree in a car, a riding time in a car, a probability of occurrence of a car call, and the like have been proposed.

Furthermore, recently, a method has been proposed in which a fuzzy amount is used as an evaluation index, and a rule group in which an appropriate allocation method is described in the IF-THEN format is used to select and assign an optimum car from the degree of conformity to the rule group. ing.

However, when trying to execute detailed group management control to respond to complicated traffic patterns and ever-changing traffic demands, the evaluation formulas and allocation evaluation rules for obtaining the above evaluation values become more and more complicated. To go. Therefore,
When trying to improve the accuracy of various prediction values used as evaluation elements, the calculation formula of the prediction value becomes complicated. In addition, developing a new arithmetic expression aiming at optimal group management control is a difficult task because human capabilities are limited.
Furthermore, on the other hand, the calculation time for performing complicated calculations is increased, and it becomes very difficult to fulfill the main function of determining and predicting the assigned car at the same time as the hall call registration.

In order to solve such a problem, for example, Japanese Patent Application Laid-Open
As described in Japanese Patent Application No. 275381, a group management control device that selects a car to be assigned to a hall call based on an operation using a neural network corresponding to neurons of the human brain has also been proposed. According to this group management control, there is no need for a human to consider an allocation algorithm, and a judgment system that determines an optimal allocated car in response to various traffic conditions can be automatically generated. However, in the case of this publication, the calculation of the evaluation value of the assigned car is considered, and the improvement of the calculation accuracy of the expected arrival time and the calculation accuracy of the expected congestion degree in the car is not considered.

Here, in the elevator control, if attention is paid to the fact that the neural network expresses the causal relationship between the traffic condition (input data) and the evaluation value (output data) of the assigned car in the network, the estimated arrival time It can be seen that the present invention can be applied to calculation, calculation of unexpected probability and fullness probability, prediction of the degree of congestion in the car, prediction of the riding time in the car, and the like.

Hereinafter, an elevator control device in a case where a neural network is applied to the calculation of the estimated arrival time will be described with reference to FIGS. 6 to 10.

In the functional block diagram of FIG. 6, the group management device (10)
Means the following means (10A) to (10D), (10F) and (10G)
And a plurality (for example, for Unit 1 and Unit 2)
It controls the car control devices (11) and (12).

The hall call registration means (10A) registers and cancels the hall calls (up and down hall calls) of each floor and the time elapsed since the hall call was registered (ie,
Duration).

The allocating means (10B) for selecting and allocating the best car for servicing the hall call, for example, predicts and calculates the waiting time until each car responds to the floor call on each floor, and calculates their squared values. Is assigned to a car that minimizes the sum of

The data conversion means (10C) converts input data such as a car position, a driving direction, and a call to be answered (a car call or an assigned landing call) into input data of a neural network. And a conversion means for converting output data of the neural network (data corresponding to the expected arrival time) into a form usable for an operation for a predetermined control purpose (for example, calculation of a predicted waiting time). .

The expected arrival time calculation means (10D) for calculating the expected arrival time of each car according to the time zone includes an expected arrival time calculation unit (described later) formed by a neural network.

The learning data creating means (10F) includes an estimated arrival time of each car and input data (traffic state data) at that time,
Then, it stores actual measurement data (teacher data) relating to the arrival time of each car, and outputs them as learning data.

The correction means (10G) learns and corrects the configuration of the neural network in the expected arrival time calculation means (10D) using the learning data.

Car control devices for Unit 1 and Unit 2 (11) and (1
2) have the same configuration. For example, the car control device (11) for the first car is composed of well-known means (11A) to (11E).

The hall call canceling means (11A) outputs a hall call cancel signal for a hall call on each floor. The car call registration means (11B) registers a car call of each floor. The arrival forecast light control means (11C) is provided with an arrival forecast light (not shown) on each floor.
The lighting of is controlled. The operation control means (11D) controls the running and stopping of the car in order to determine the operating direction of the car and to respond to the car call and the assigned hall call.
The door control means (11E) controls opening and closing of the door at the entrance of the car.

In the block diagram of FIG. 7, the group management device (10)
It consists of a well-known microcomputer, MPU (micro processing unit) or CPU (101) and MPU (101)
And an input circuit (104) and an output circuit (105) connected to the MPU (101).

The input circuit (104) includes a hall button signal (14) from a hall button on each floor and one of the signals from the car control devices (11) and (12).
The status signals of the first and second units are input. Also, from the output circuit (105), a hall button light signal (15) to a hall button light built in each hall button, and car control devices (11) and (12)
And a command signal to the controller.

FIG. 8 is a functional block diagram specifically showing the relationship between the data conversion means (10C) and the expected arrival time calculation means (10D) in FIG. 6 through a network.

In the figure, an input data conversion means, ie, an input data conversion subunit (10CA), and an output data conversion means, ie, an output data conversion subunit (10CB), constitute a data conversion means (10C). Input data conversion subunit (10CA) and output data conversion subunit (10CB)
Estimated arrival time calculation unit (10DA) inserted between
Is composed of a neural network and constitutes a prediction calculation subroutine used by the expected arrival time calculation means (10D).

Input data conversion sub-unit (10CA)
Traffic state data such as direction of travel, calls to be answered (ie, car calls and assigned hall calls), and statistical characteristics of traffic flow (5 minutes passengers, 5 minutes drop-offs) are stored in a neural network (10DA). Convert to a form that can be used as input data.

The output data conversion subunit (10CB) converts the output data (data corresponding to the expected arrival time) of the neural network (10DA) into a form that can be used for calculating the evaluation value of the hall call assignment operation.

The expected arrival time calculation unit (10DA) composed of a neural network includes an input layer (10DA1) that takes in input data from the input data conversion subunit (10CA) and an output layer (data) that outputs data corresponding to the expected arrival time. 10DA
3) and an intermediate layer (10DA2) between the input layer (10DA1) and the output layer (10DA3), in which a weight coefficient is set.

These layers (10DA1) to (10DA3) are connected to each other via a network, and each of the layers (nodal
e).

Here, if the number of nodes of the input layer (10DA1), the middle layer (10DA2) and the output layer (10DA3) is N1, N2, and N3, respectively, the number of nodes N3 of the output layer (10DA3) is N3 = 2. (FL-1) where FL is the number of floors of the building, and the number of nodes N1 and N2 of the input layer (10DA1) and the intermediate layer (10DA2) are the number of floors FL of the building and the input data to be used, respectively. And the number of cars.

If the variables i, j, k are i = 1, 2,..., N1 j = 1, 2,..., N2 k = 1, 2,..., N3, the i-th node of the input layer (10DA1) Are input values and output values of xa1 (i) and ya1 (i), and input values and output values of the j-th node of the intermediate layer (10DA2) are xa2 ((j) and ya2
(J) The input value and output value of the k-th node of the output layer (10DA3) are represented by xa3 (k) and ya3 (k).

Also, the i-th node of the input layer (10DA1) and the intermediate layer (10DA2)
The weight coefficient between the j-th node of the output layer (10DA3) and the weight coefficient between the j-th node of the intermediate layer (10DA2) and the k-th node of the output layer (10DA3) is wa2 (j, k). Then, the relation between the input value and the output value of each node is as follows: ya1 (i) = 1 / [1 + exp {-xa1 (i)}] xa2 (j) = {wa1 (i, j) * ya1 (I)｝ (sum expression by i = 1 to N1) ya2 (j) = 1 / [1 + exp ｛-xa2 (j)｝] xa3 (k) = Σ ｛wa2 (j, k) × ya2 (j )｝ (Sum expression by j = 1 to N2) ya3 (k) = 1 / [1 + exp {−xa3 (k)}]

The neural network (10DA) is connected to a learning data creation unit (not shown) used in the learning data creation means (10F) and a correction unit (not shown) used in the correction means (10G). Weight coefficient wa1 (i, j)
And wa2 (j, k) are appropriately modified.

Hereinafter, the operation of calculating the expected arrival time of the elevator control device shown in FIGS. 6 to 10 will be described with reference to the flowchart of FIG.

First, by the input data conversion program (step 91), of the input traffic condition data, data (car position, operation direction, car call, assigned hall call) regarding the car from which the expected arrival time is to be calculated, Data showing the statistical characteristics of the traffic flow in the country (number of passengers for 5 minutes,
5 minutes), and converts them as input data xa1 (1) to xa1 (N1) for each node of the input layer (10DA1) of the expected arrival time calculation unit (10DA).

Here, the number of floors FL of the building is assumed to be the 12th floor, and f = 1, 2,..., 11 represent the ascending landings on the 1, 2,. , 13,…, 22 are 12,11,…, 2 respectively
Assuming that the landing at the floor in the down direction is represented, for example, the car state of "floor at the car position floor and up in the operation direction" is xa1 (f) = 1 xa1 (i) = 0 (i = 1,2, .., 22, i） f), and is represented as a value normalized to a value of 0 to 1.

In addition, car calls xa1 (23) to xa1 (34) on the 1st to 12th floors
Is represented as "1" if registered, and "0" if not registered. The assigned floor calls xa1 (35) to xa1 (45) in the first to eleventh floors in the upward direction are allocated. If it is not allocated, it is represented by "0", and if it is not allocated, it is represented by "0", and the allocated floor calls xa1 (46) to xa1 (5
6) is represented by "1" if assigned, and "0" otherwise.

In addition, the number of passengers xa1 for 5 minutes in the upward direction from the 1st floor to the 11th floor
(57) to xa1 (67) are obtained by dividing the number of passengers per 5 minutes obtained from the past traffic volume statistics by the maximum possible value NNmax (for example, 100) to obtain a value of 0 to 1. Normalize. Similarly, the number of passengers xa1 (68) to xa1 (78) in the descending direction from the 12th floor to the second floor is xa1 (79) to xa1 (89) in the upward direction from the 1st floor to the 11th floor for 5 minutes. , And 12
The number of people getting off in the down direction from the first floor to the second floor for 5 minutes xa1 (90)-xa1
(100) is also normalized by dividing by the maximum value NNmax.

Note that the method of normalizing the input data is not limited to the above method, and the car position and the operation direction can be separately represented. For example, when the car position floor is f,
Input value xa1 (1) of the first node representing the car position floor
, Xa1 (1) = f / FL, and the input value xa1 of the second node representing the car operation direction
(2), the upward direction is “+1”, the downward direction is “−1”,
No direction may be represented as “0”.

When the input data for the input layer (10DA1) is set in step 91 in this way, the following steps 92 to 96 perform a network operation for estimating the arrival time when a new hall call C is temporarily assigned to the first car. I do.

First, the output value ya1 (i) of the input layer (10DA1) is calculated from the equation using the input data xa1 (i) (step 92).

Subsequently, the output value ya1 (i) obtained by the equation is multiplied by a weighting factor wa1 (i, j), and the sum of i = 1 to N1 is summed up. From the equation, the input value of the intermediate layer (10DA2) is obtained. xa2 (j)
Is calculated (step 93).

Subsequently, using the input value xa2 (j) obtained by the equation,
The output value ya2 (j) of the intermediate layer (10DA2) is calculated from the equation (step 94).

Subsequently, the output value ya2 (j) obtained by the equation is multiplied by a weighting coefficient wa2 (j, k), and the sum of j = 1 to N2 is summed up. From the equation, the input value of the output layer (10DA3) is obtained. xa3 (k)
Is calculated (step 95).

Then, using the input value xa3 (k) obtained by the equation,
The output value ya3 (k) of the output layer (10DA3) is calculated from the equation (step 96).

As described above, when the network calculation of the estimated arrival time is completed, the output data conversion subunit (10
CB), the form of the output values ya3 (1) to ya3 (N3) is converted to determine the final expected arrival time (step 97).

At this time, each node of the output layer (10DA3) corresponds to a landing for each direction, and the output value ya3 of the first to eleventh nodes.
(1) to ya3 (11) are used to determine the calculated values of the expected arrival times of the upbound landings on the 1,2,..., 11th floor, respectively.
The output values ya3 (12) to ya3 (22) of the twelfth to twenty-second nodes are used to determine the calculated value of the expected arrival time of the downbound landing, respectively.

That is, the output value ya3 (k) of the k-th node is converted into the estimated arrival time T (k) of the landing k, and the estimated arrival time T (k)
(K) is expressed as: T (k) = ya3 (k) × NTmax. However, NTmax is a constant value representing the maximum value that the expected arrival time can take. Here, since the output value ya3 (k) of the k-th node is normalized to the range of 0 to 1, by multiplying the maximum value NTmax as in the equation, the expected arrival time T (k) becomes It is converted so that it can be used for the evaluation value calculation of the hall call assignment.

Thus, the arrival time prediction program (steps 91 to 91)
In 97), the causal relationship between the traffic condition and the expected arrival time is expressed in a network, and the traffic condition data is taken into the neural network, whereby the expected arrival time can be calculated with high accuracy. In addition, if the assigned car for the hall call is selected based on the estimated arrival time, the waiting time for the hall call can be reduced.

Furthermore, the calculation accuracy of the network is determined by the weighting coefficients wa1 (i, j) and wa2 connecting the nodes of the neural network (10DA).
Since it changes according to (j, k), an appropriate expected arrival time can be determined by appropriately changing and correcting the weighting coefficients wa1 (i, j) and wa2 (j, k) by learning.

The learning in this case (modification of the network) is efficiently performed using the back propagation method. The backpropagation method is based on network output data,
This is a method of correcting a weight coefficient connecting a network using an error with desired output data (teacher data) created from measured data, control target values, and the like.

That is, learning data creation means (10F) (see FIG. 6)
The learning data creation unit in (a), when a predetermined time (for example, when a hall call is allocated) during operation of the elevator, ya3 (1) representing the estimated arrival time of each hall
~ Ya3 (N3) and traffic condition data xa1 (1) at that time
Xa1 (N1) is stored as a part of the learning data. Then, the elapsed time until the car stops or passes through the landing is counted, and the actual arrival time is stored as a part of the learning data. This is the original teacher data, and is represented by the estimated arrival time TA (k) (k = 1, 2,..., N3).
Such a set of learning data is sequentially stored each time a predetermined condition is satisfied.

Next, when the correction unit in the correction means (10G) detects that it is time to perform network correction, the correction unit (10G) checks the network in the expected arrival time calculation unit (10DA) based on the learning data in FIG. Is corrected according to the flowchart of FIG.

First, it is determined whether it is time to modify the network (step 111). If it is time to modify the network, the following steps 112 to 118 are executed.

Here, the time when the number m of the currently stored learning data sets becomes S (for example, 500) or more is defined as the network correction time. In addition, the judgment reference number S of the learning data is the number of elevators installed, the number of floors FL of the building, and
It can be set arbitrarily according to the scale of the network such as the number of hall calls.

If it is determined in step 111 that the number m of the learning data sets is S or more, the counter number n of the learning data is initialized to “1” (step 112), and then the n-th learning data From the actual arrival time TA (1) to TA (N3)
And the values of the nodes corresponding to these landings, that is, the teacher data da (k) (k = 1, 2,..., N3) are obtained from da (k) = TA (k) / NTmax. 113).

Next, the error Ea between the output values ya3 (1) to ya3 (N3) of the output layer (10DA3) extracted from the n-th learning data and the teacher data da (1) to da (N3) is calculated. And the sum of k = 1 to N3 is obtained from Ea = {[{da (k) -ya3 (k)} ² ] / 2 (k = 1 to N3). Then, using the error Ea obtained by the equation,
Weight coefficient wa between the hidden layer (10DA2) and the output layer (10DA3)
2 (j, k) (j = 1, 2,..., N2, k = 1, 2,..., N3) are modified as follows (step 114).

First, the error Ea of the equation is differentiated from wa2 (j, k), and rearranged using the above-described equations (1) to (2), the change amount Δwa2 (j, k) of the weighting coefficient wa2 (j, k) becomes Δwa2 (j , k) = − α ｛∂Ea / {wa2 (j, k)} = − α · δa2 (k) · ya2 (j). Here, α is a parameter representing the learning speed, and can be selected to any value within the range of 0 to 1.
In the equation, δa2 (k) = {ya3 (k) −da (k)} ya3 (k) ｛1−ya3 (k)}. Thus, the change amount Δwa2 of the weight coefficient wa2 (j, k)
When (j, k) is calculated, the weighting coefficient wa2 is calculated by the following equation.
The correction of (j, k) is performed.

wa2 (j, k) ← wa2 (j, k) + Δwa2 (j, k) Also, similarly, the weighting coefficient wa1 (i, j) (i = 2) between the input layer (10DA1) and the intermediate layer (10DA2). 1,2,…, N1, j = 1,2,…, N
2) is modified according to the following formula and formula (step 115).

First, the variation Δwa1 (i, j) of the weight coefficient wa1 (i, j)
Is obtained from Δwa1 (i, j) = − α · δa1 (j) · ya1 (i). However, in the equation, δa1 (j) is the following summation formula with k = 1 to N3, δa1 (j) = Σ ｛δa2 (k) · wa2 (j, k) · ya2 (j) × [1-ya2 (J)]｝. Using the variation Δwa1 (i, j) obtained by the equation, the weight coefficient wa1 (i, j) is corrected as in the following equation.

wa1 (i, j) → wa1 (i, j) + Δwa1 (i, j) In this manner, the correction step using the n-th learning data
When steps 113 to 115 are performed, the learning data number n is incremented (step 116), and it is determined in step 117 that the correction has been completed for all the learning data (n ≧ m).
Are repeated until the following.

When all the learning data have been corrected, the weight coefficients wa1 (i, j) and wa2 (j,
k) is registered in the expected arrival time calculation means (10D) (step 118).

At this time, all the learning data used for the correction are cleared so that the latest learning data can be stored again, and the learning data number m is initialized to “1”. Thus,
Neural network (10DA) network modification (learning)
To end.

As described above, the causal relationship between the traffic state when the hall call is registered and the estimated arrival time is expressed in the network, and the network can be corrected by learning the actual data. As a result, it is possible to automatically cope with a change in the traffic flow in the building, and it is possible to perform a flexible, precise and accurate calculation of the estimated time of arrival.

However, since a single neural network (10DA) is used for the prediction calculation, if a large amount of input data is taken in to improve the calculation accuracy, it takes a long calculation time, and the generation timing of the output data, that is, the expected arrival time is reduced. In the end, it will not be possible to determine the appropriate assigned car.

[Problem to be Solved by the Invention] As described above, the conventional elevator control apparatus calculates output data by using a neural network (10DA) by flexible prediction according to actual traffic conditions, The appropriate car has been assigned based on that. However, since the characteristics of the traffic flow in an elevator are constantly changing even during the day, an accurate estimated time of arrival can be obtained only by using the car position, the operation direction, and the call to be answered as input data. Cannot be calculated.

Therefore, data representing characteristics of traffic flow, for example,
It is conceivable to use the traffic amount (the number of passengers getting on and off, the number of hall calls, the number of car calls, etc.) statistically used in the past as input data. According to this, flexible calculation according to various traffic conditions that can occur throughout the day can be realized using one neural network in the expected arrival time calculation means (10D).
However, as the input data increases, it takes a long time to calculate the estimated arrival time, so that it becomes very difficult to perform the function of determining and predicting the assigned car at the same time as the hall call registration. Also, the weighting factor wa1 (i,
In order to correct j) and wa2 (j, k), there is a problem that much learning data (teacher data) and a learning period are required.

Such a problem is a common problem when the neural network is applied to the calculation of the assigned car and the calculation of various predicted values.

Further, as described in JP-A-55-48174,
A group management control device that includes an evaluation formula parameter storage unit and a traffic pattern setting unit in order to allocate a car having an optimum evaluation value and changes the evaluation formula (weight coefficient) according to the traffic pattern has also been proposed. Is a table value written in advance in a data area (PROM) in the computer, the weighting coefficient is merely changed by being selected from predetermined table values. High accuracy cannot be achieved.

The present invention has been made in order to solve the above-described problems, and an elevator control capable of performing a calculation according to a predetermined control purpose in a short time and with high accuracy even for various traffic flows. The aim is to obtain a device.

Still another object of the present invention is to provide an elevator control apparatus capable of correcting a network with a small amount of learning data and a short learning period.

[Means for Solving the Problems] An elevator control device according to the present invention includes
A plurality of determining means for providing a plurality of traffic patterns corresponding to a plurality of traffic patterns classified according to the characteristics of traffic flow in the building, and determining which of the traffic patterns the current elevator traffic state corresponds to; Switching means for selecting only one of the calculating means corresponding to the determination result of the determining means, and controlling the car based on the output data of the calculating means selected by the switching means. It is.

Also, in the elevator control device according to another invention of the present invention, the learning data creating means creates learning data for each traffic pattern based on the determination result of the determining means, and the correcting means creates learning data for each traffic pattern. The weight coefficient of the arithmetic means corresponding to each traffic pattern is corrected using the data for traffic, and the car is controlled based on the output data of the arithmetic means selected by the switching means.

[Operation] In the present invention, only one computing means corresponding to the current traffic condition is selected from a plurality of computing means provided corresponding to the traffic patterns classified according to the characteristics of the traffic flow. Then, the elevator is controlled in accordance with a predetermined purpose based on the output data of the selected calculating means.

Further, in another aspect of the present invention, learning data is created for each traffic pattern, and the weighting factors of the arithmetic means corresponding to each traffic pattern are corrected based on the learning data.

Embodiment An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a functional block diagram showing the overall configuration of an embodiment of the present invention, wherein (10), (10A) to (10D) (10G), (1)
1), (11A) to (11E) and (12) are the same as those shown in FIG. The schematic configuration of the elevator control device shown in FIG. 1 is as shown in FIG.

In FIG. 1, the expected arrival time calculation means (10D)
Expected arrival time calculation means (10D1) for normal hours, expected arrival time calculation means (10D2) for work hours, expected arrival time calculation means (10D3) for leaving hours, and lunch time It includes expected arrival time calculation means (10D4) and expected arrival time calculation means (10D5) for off-peak hours. The network configuration of each calculation means (10D1) to (10D5) is the same as in FIG. .

However, in the neural network in each of the calculation means (10D1) to (10D5), the traffic volume data used in FIG. 8 (the number of passengers for 5 minutes and the number of passengers for 5 minutes) is not used. Therefore,
In this case, if the number of floors of the building is 12F, the input layer (10DA
The number of nodes N1 'in 1) is set to 56. This is because the number of passengers and the number of passengers in the up and down directions for each landing are deleted from the input data.
This is because the value is 44 less than (= 100).
The number of nodes N2 'in the intermediate layer (10DA2) is set to a value smaller than N2 as the number of input data decreases.

A group management device (10) for determining which traffic pattern is the current traffic condition of the elevator (10E)
And a switching means (10H) for selecting only one of a plurality of expected arrival time calculating means (10D1) to (10D5) according to the result of the determination by the determining means (10E). In addition, a plurality of traffic patterns representing the traffic condition of the elevator include:
It is assumed that a distinction between a plurality of time zones is included.

FIG. 2 is a flowchart schematically showing a group management program stored in the group management device (10), and FIG.
FIG. 4 is a flowchart specifically showing the arrival time prediction program in FIG. 4, FIG. 4 is a flowchart specifically showing the learning data creation program in FIG. 1, and FIG. 5 is a concrete example of the correction program in FIG. FIG. 4 is a flowchart diagram schematically shown.

Hereinafter, the group management operation of the embodiment of the present invention shown in FIG. 1 will be described with reference to FIG.

First, the group management device (10) fetches a hall button signal (14) and status signals from the car control devices (11) and (12) according to a well-known input program (step 31). The state signal input here includes a car position, a traveling direction, a stop or traveling state, a door open / close state, a car load, a car call, a hall call cancellation signal, and the like.

Next, the well-known hall call registration program (step 32)
, The registration or cancellation of the hall call and the lighting or extinguishing of the hall button light are determined, and the duration of the hall call is calculated.

Subsequently, it is determined according to a known determination program (step 33) which time zone the current elevator traffic state is.

For example, according to the output of a clock (not shown) built in the group management device (10), the work time (8:30 to 9:10), the leave time (17:00 to 17:30), the lunch time (11: 50-13: 1
0), off-peak hours (0: 00-8: 30, and 19: 00-24: 0)
0) or a normal time zone (a time zone other than the above time zones).

Next, it is determined whether or not a new hall call C has been registered (step 34). If a newly registered hall call C is detected, the following steps 35 to 39 determine the hall call C.
The calculating a first car and a waiting time evaluation value W ₁ when the assigned respectively to Unit 2 and W ₂ (see JP-B-58-48464). Then, select the car evaluation value W ₁ or W ₂ is minimized as a regular assigned car for the assigned car, setting the allocation command and forecasts command corresponding to the landing call C.

That is, first, a new landing call C is set to 1 by the arrival time prediction program for the temporary allocation for the first car (step 35).
The expected arrival time Ta1 (k) (k = 1, 2,..., N3) of each of the first landings at each landing when temporarily assigned to the first landing is calculated.

Similarly, the expected arrival time Ta2 for each landing of the second car when the hall call C is temporarily allocated to the second car by the arrival time prediction program at the time of the temporary assignment of the second car (step 36).
(K) is calculated.

In addition, the arrival time prediction program (steps 37 and 38) at the time of non-temporary assignment when the new hall call C is ignored and not assigned to either of the first and second units is executed, and each of the first and second units is executed. The expected arrival times Tb1 (k) and Tb2 (k) for the landing are calculated.

Hereinafter, the calculation operation of the arrival time prediction program (35) at the time of provisional assignment for the first car will be specifically described with reference to FIGS. 3 and 8.

First, the hall call C is temporarily allocated to the first car, and allocated call data for input to the input data conversion subunit (10CA) is created (step 51).

Next, based on the determination result of the determination program (step 33) in the determination means (10E), the switching means (10H) calculates the expected arrival time corresponding to each of the expected arrival time calculation means (10D1) to (10D5). Only one program is selected from the programs (steps 56 to 60) (steps 52 to 55).

Here, only the normal time zone expected arrival time calculation program (step 56) is specifically shown, but each expected arrival time calculation program (steps 57 to 60) is also referred to as step 56.
It is composed of the same arithmetic program.

For example, when it is determined by the determination program (step 33) in the determination means (10E) that the time period is a normal time zone, the work time zone determination step 52, the leaving time zone determination step 53, the lunch time zone determination step 54, the idle time The program proceeds to a normal time zone expected arrival time calculation program (step 56) via the band determination step 55, and the same calculation program as that of FIG. 9 is executed.

Steps 561 to 567 in step 56 correspond to steps 91 to 97 in FIG. 9, respectively. Also, weighting coefficients wa1 (i, j) (i = 1,2,...,
N1 ′, j = 1, 2,..., N2 ′) and wa2 (j, k) are set to values for normal time zones.

In the input data conversion program (step 561), 1
The data such as the car position, the operation direction, the car call of the car, the assigned call after the provisional assignment, etc. are extracted and converted into input data xa1 (1) to xa1 (i) for network operation. Here, i = 1, 2,..., N1 ′ (N1 ′ <N1). Hereinafter, similarly to the above-described steps 92 to 97, the network operation in steps 562 to 567 is executed, and finally, the estimated arrival time Ta1
(K) (k = 1, 2,..., N3) is set.

On the other hand, when the time zone corresponding to each of the determination steps 52 to 55 is determined, the respective expected arrival time calculation programs (steps 57 to 60) are similarly executed. The weighting coefficients wa1 (i, j) and wa2 used in each network
(J, k) (i = 1,2, ..., N1 ', j = 1,2, ..., N2', k = 1,
2, ..., N3) are set to values corresponding to the respective time zones.

As described above, in each of the arrival time prediction programs 35 to 38, since the traffic data is removed from the input data, the number of input data is reduced from 100 to 58 and the intermediate layer (10D
A2) The number of nodes can be reduced. Also, from the estimated arrival time calculation program (steps 56 to 60) composed of a plurality of neural networks provided corresponding to the time zones,
Only one calculation program corresponding to the current traffic condition is selected, and estimated arrival times Ta1 (k), T1
Since a2 (k), Tb1 (k) and Tb2 (k) are calculated, the expected arrival time can be calculated in a short time and with high accuracy.

Estimated arrival time obtained in this manner is the allocation program (step 39), it is used in the calculation of the waiting time evaluation value W ₁ and W _2.

Next, according to the output program (step 40), the hall button light signal (15) set as described above is sent to the hall, and the assignment signal and the forecast signal are transmitted to the car controller (1).
Send to 1) and (12).

On the other hand, the learning data creation program (step 41) stores, as input data, the traffic condition data after conversion, the estimated arrival time of each landing, and the actual measurement data of the arrival time of each car thereafter, and learns them. Output as data for use. In the correction program (step 42), the weight coefficient of the network in the expected arrival time calculation means (10D) is corrected using the learning data.

Next, referring to FIGS. 4 and 5, when the learning data creation program (step 41) and the correction program (step 42) are executed by the learning data creation means (10F) and the correction means (10G). Next, another embodiment of the present invention will be described.

In FIG. 4 showing the learning data creation program (step 41) in detail, first, a permission to create new learning data is generated (set), and whether a new hall call C has been allocated is performed immediately. It is determined whether or not it is (step 61).

If the permission to create learning data is set,
If the hall call C has been allocated, the traffic condition data xa1 (1) to xa1 (N1 ') of the allocated car at the time of the allocation.
And output data ya3 () to ya3 (N3) corresponding to the estimated arrival time of each landing at this time are stored as part of the m-th learning data (teacher data) (step 62).

Subsequently, the permission to create new learning data is reset, and the actual arrival time actual measurement command is set to start counting the actual arrival time (step 63).

Thereby, in step 61 of the next operation cycle,
Since it is determined that the permission to create new learning data has not been set, the process proceeds to step 64, and it is determined whether an actual measurement instruction for the arrival time has been set. At this time,
Since the actual measurement command is set in step 63,
Further, in step 65, it is determined whether or not the assigned car has stopped in response to the hall call C.

If it is determined that the stop of the hall call C to the hall is detected in the calculation cycle several times later, the process proceeds to step 66, and the actual arrival time at this time is stored as a part of the m-th learning data. This is the original teacher data, and is expressed as the arrival time TA (C) of the hall at the hall call C.

Subsequently, in step 67, an actual measurement command of the actual arrival time is issued to terminate the counting of the actual arrival time, the number m of the learning data is incremented, and the permission to create new learning data is set again.

In this way, according to the time at which the hall calls are allocated, whether the hall call ratio has been performed and the learning data relating to the new hall call C at that time are repeatedly created and stored.

Next, the correcting means (10G) corrects the network of the neural network (10DA) by using the learning data according to the correcting program (step 42) in FIG. Hereinafter, this correction operation will be described in detail with reference to FIG.

First, it is determined whether it is time to modify the network (step 71). If it is time to modify the network, the following steps 72 to 84 are executed.

Here, the time when the number m of the currently stored learning data sets becomes S (for example, 400) or more is defined as the network correction time. In addition, the judgment reference number S of the learning data is the number of elevators installed, the number of floors FL of the building, and
It can be set arbitrarily according to the scale of the network such as the number of hall calls.

If it is determined in step 71 that the number m of the learning data sets is S or more, the learning data counter number n
Is initially set to "1" (step 72), it is determined which time zone the n-th learning data corresponds to (steps 73 to 76), and a plurality of correction programs (steps 77 to 81) are determined. ), Select only one correction program to be used. Here, only the correction program (step 77) for the weight coefficient for the normal time zone is specifically shown,
Each correction program (steps 78 to 81) is also a program similar to step 77.

For example, when it is determined that the n-th learning data is data for the normal time zone, the process proceeds to the correction step 77 via each of the determination steps 73 to 76, and the correction program for the weight coefficient for the normal time zone is executed. Selected and executed. Each of steps 771 to 773 in step 77 corresponds to step 11 in FIG.
3 to 115 respectively.

First, from the n-th learning data, the actual arrival time TA
After extracting (C) and obtaining teacher data da (c) (step 771), the middle layer (10DA2) and the output layer (10DA3)
The weighting coefficient wa2 (j, C) (j = 1,2,..., N2 ') between the input layer (10DA1) and the intermediate layer (10DA2) is corrected. Weight coefficient wa1 (i, j) (i = 1,2,
.., N1 ') are corrected (step 773).

On the other hand, if the n-th learning data is data for the work time zone, the program for correcting the weight coefficient for the work time zone (step 78) is selected in step 73 of determining whether the learning data is for the work time zone. Then, the weighting factor in the expected arrival time calculation program for a work time zone (step 57) is corrected. Similarly, if the n-th learning data is data for a leaving time zone, a correction program (step 79) for a weight coefficient for a leaving time zone is selected in a decision step 74, and if it is a lunch time period, a decision step is performed. The program for correcting the weight coefficient for the lunch time zone is selected according to 75 (step 80), and if it is the off-hours, the correction program for the weight coefficient for off-hours time is selected at the determination step 76 (step 81).
Each weighting factor is modified. The specific correction procedure is the same as described above, and will not be described in detail here.

As described above, when the correction by the n-th learning data (steps 73 to 81) is completed, the number n of the learning data is incremented (step 82), and the correction is completed for all the learning data in step 83. (N
Until ≧ m), the processing of steps 73 to 82 is repeated.

When all the learning data have been corrected, the weight coefficients wa1 (i, j) and wa2 (j,
k) is registered in the expected arrival time calculation means (10D) (step 84).

At this time, all the learning data used for the correction are cleared so that the latest learning data can be stored again, and the learning data number m is initialized to “1”. Thus,
Correction (learning) of each neural network (10DA) for each time zone
To end.

As described above, since the learning data is created for each time zone and the network of the expected arrival time calculating means is modified for each time zone, the expected arrival time calculating means (10D) composed of one neural network is used. The correction can be performed with less learning data and in a shorter learning period than when the network is corrected. As a result, the estimated arrival time can be accurately calculated in a short time for various traffic flows.

In general, when the number of input points (data amount) increases, the calculation time further increases. However, the neural network is constructed for each of a plurality of traffic patterns (time periods such as normal times, going to work and leaving work or traffic conditions). By providing the data representing the characteristics of the traffic flow from the input data of the neural network, the calculation time can be greatly reduced. The learning is also performed for each traffic pattern, and the control index includes predicted values such as an arrival time, a predicted waiting time, a degree of congestion in a car, and a possibility of a missed forecast, as well as an assigned car determination.

In other words, by performing prediction calculations such as expected arrival time, forecast deviation, car deviation, and congestion degree in the car for each time zone or each traffic condition, it is not necessary to input the characteristic points (statistics) of the traffic flow. Time and learning time can be significantly reduced. This has a special effect in an elevator apparatus in which the traffic flow changes for each time zone, and accurate prediction results and calculation results are required for each time zone in a short time.

In the above embodiment, the input data conversion means converts the car position, the driving direction, and the call to be answered as input data. However, the traffic condition data used as input data is limited to these. It will not be done. For example, car state (during deceleration, door opening operation, door opening, door closing operation, door closing standby, running, etc.), hall call duration, car call duration, car load, group The number of managed cars can be used as input data. By using not only the current traffic condition data but also the near future traffic condition data (history of car movement and call response status) as input data,
More accurate calculation of the estimated arrival time can be performed.

The determining means (10E) determines which time zone the current elevator traffic state is based on whether or not the output of the clock has reached a predetermined time zone. However, it is not limited to this. For example, the determination can be made in consideration of the number of passengers (car load) of a car departing from the congested floor or the number of passengers on the congested floor that exceeds a predetermined value. In this case, the type of the time zone is appropriately set according to the traffic condition of the building. Regardless of the time zone, a plurality of typical traffic flow patterns (for example, a pattern in which traffic is skewed in the upward direction, a pattern in which the traffic is skewed in the downward direction, etc.) are prepared, and Based on the actual measurement value of the traffic data (for example, 5 minutes), it may be determined which traffic pattern the current traffic is close to, and the closest traffic pattern may be selected.

The learning data creating means (10F) stores, when the hall call is assigned, the estimated arrival time of the assigned car at the hall call, the input data at that time, and the time zone, Thereafter, the elapsed time until the assigned car stops at the landing call hall is counted and stored as the actual arrival time, and the stored time zone, input data, estimated arrival time,
Although the actual arrival time is output as a set of learning data, the time at which the learning data is created is not limited to this. For example, the time when the time elapsed from the previous storage of the input data exceeds a predetermined time (for example, one minute) may be set as the learning data generation time, and the learning data generation time may be periodically (for example, every one minute). It may be time. Further, the more learning data is collected under various conditions, the more the learning conditions are improved. For example, when the car is stopped on a predetermined floor, or in a predetermined state (during deceleration, stopping, etc.). It is also possible to determine in advance a typical state that can be considered, and to create learning data when the state is detected. Further, the way of storing the learning data is not limited to this, and the learning data may be stored in the storage area for each time zone in a distinguished manner. In this case, the amount of data that must be stored as learning data can be reduced.

The correction means (10G) for correcting the weight coefficient of the expected arrival time calculation means (10D) corrects the weight coefficient every time the total number m of the stored learning data reaches the predetermined value S. The modification time is not limited to this. For example, at a predetermined time (for example, every hour), the weighting factor may be corrected based on the learning data stored up to that time, and traffic is reduced, and the calculation frequency of the estimated arrival time is reduced. Sometimes, the weighting factor may be modified. Also, the judgment is not made based on the total number m of the learning data, but the number of learning data mA, mB,..., ME for each time zone is separately counted, and each predetermined value SA, SB,
.., Each time the SE is reached, the weight coefficient of the corresponding time zone may be corrected.

Also, in the set of learning data created by the learning data creating means (10F), only the expected arrival time and the actual arrival time for one hall (new hall call C) are stored in addition to the input data. When the weighting coefficient is corrected by the correcting means (10G), only the weighting coefficient related to the learning data (teacher data) is corrected, but the learning method is not limited to this. For example, the estimated arrival time for all landings and the actual arrival time measured during the operation of the car are stored, and only the weight coefficient related to the teacher data of the landing where the actual arrival time exists is corrected. You may do so. In this case, the amount of data that must be stored as learning data can be reduced. In addition, the landing where the actual arrival time could not be measured is, for example, when the direction of the car is reversed on the floor on the way,
This corresponds to a landing farther than the inverted floor. If the car becomes an empty car (a car with no assigned call) on the middle floor, the landing farther than the empty floor and the input data This corresponds to a landing behind the floor at the car position at the time of storage (for example, a landing below the current position during upward operation).

Further, in the above embodiment, in order to predict the movement of the car in the near future, the neural network is used for calculating the estimated arrival time. However, the allocation of hall calls and other prediction items used for group management control are performed. The same can be applied to calculations. For example, the present invention can be applied to a forecast deviation probability, a fullness probability, a prediction of a car load on each floor, a prediction of a car call occurrence, and the like.

Further, the correction step of the weight coefficient is performed a plurality of times (for example, 50
The weighting factor may be converged so that a desired approximate output is obtained by repeating 500 times for 0 data.

[Effects of the Invention] As described above, according to the present invention, a plurality of arithmetic means are provided corresponding to a plurality of traffic patterns classified according to the characteristics of the traffic flow in the building, and the current elevator Determining means for determining which of the traffic patterns the traffic state corresponds to, and switching means for selecting only one of the calculating means corresponding to the determination result of the determining means,
Since the car is controlled based on the output data of the calculation means selected by the switching means, an elevator which can execute a calculation approximate to a predetermined control purpose with high accuracy in a short time even for various traffic flows. There is an effect that a control device can be obtained.

According to another aspect of the present invention, at a predetermined time during the operation of the elevator, the output data by the arithmetic means and the input data used at that time are stored, and the teacher obtained by the control result is stored. Learning data creating means for storing data and outputting the stored input data, output data, and teacher data as a set of learning data; and correcting means for correcting the weight coefficient of the arithmetic means using the learning data. The learning data creating means creates learning data for each traffic pattern based on the determination result of the determining means, and the correcting means uses the learning data for each actual traffic pattern to generate the learning data for each traffic pattern. Since the weighting factors of the calculation means corresponding to are modified, the network can be modified with a small amount of learning data and a short learning period. The effect of high accuracy of the elevator control device in response to changes in the flow can be obtained.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a functional block diagram showing an overall configuration of an embodiment of the present invention and another embodiment of the present invention. FIG. 2 shows a group management program by a group management device in FIG. FIG. 3 is a flowchart schematically showing the arrival time prediction program in FIG. 2; FIG. 4 is a flowchart specifically showing the learning data creation program in FIG. 2; FIG. 5 is a flowchart specifically showing the correction program in FIG. 2, FIG. 6 is a functional block diagram showing an example of an elevator control device which can be considered based on a conventional elevator control device, and FIG. FIG. 8 is a block diagram schematically showing a group management device, FIG. 8 is a block diagram specifically showing data conversion means and expected arrival time calculation means in FIG. 6, and FIG. 9 is a group in FIG. Calculation by management device FIG. 10 is a flowchart showing a program, and FIG. 10 is a flowchart specifically showing a correction program by the group management device in FIG. (10C) Data conversion means (10CA) Input data conversion subunit (10CB) Output data conversion subunit (10DA) Neural network (10DA1) Input layer (10DA2) Intermediate layer (10DA3) Output layer (10D) Expected arrival time calculation means (10D1) to (10D5) Multiple expected arrival time calculation means (10E) Judgment means (10F) Learning data creation means (10G)... Correction means, (10H)... Switching means wa1 (i, j), wa2 (j, k)... Weighting coefficients In the drawings, the same reference numerals indicate the same or corresponding parts.

Claims (2)

(57) [Claims] 1. An apparatus for controlling a car of an elevator installed on a plurality of floors, said input data conversion for converting traffic state data of said elevator into a form usable as input data of a neural network. Means, an input layer that captures the input data, an output layer that uses a calculation result according to a predetermined control purpose as output data, and an intermediate layer between the input layer and the output layer where a weight coefficient is set. And an output data conversion means for converting the output data into a form that can be used for the operation for the control purpose. A plurality of traffic patterns are provided corresponding to a plurality of traffic patterns classified according to the characteristics of the traffic flow. Determining means for determining which one of the patterns corresponds to, and switching means for selecting only one of the arithmetic means corresponding to the determination result of the determining means; An elevator control device, wherein the car is controlled based on output data of a calculating means. 2. An apparatus for controlling a car of an elevator installed on a plurality of floors, the input data conversion for converting the traffic condition data of the elevator into a form usable as input data of a neural network. Means, an input layer that captures the input data, an output layer that uses a calculation result according to a predetermined control purpose as output data, and an intermediate layer between the input layer and the output layer where a weight coefficient is set. Calculating means for configuring the neural network, output data converting means for converting the output data into a form usable for the operation for the control purpose, and calculating at a predetermined time during operation of the elevator. The output data by the means and the input data used at that time are stored, and the teacher data obtained by the control result is stored and stored. Said input data,
An elevator control device comprising: a learning data generating unit that outputs the output data and the teacher data as a set of learning data; and a correcting unit that corrects a weight coefficient of the arithmetic unit using the learning data. In the above, a plurality of the arithmetic means are provided in correspondence with a plurality of traffic patterns classified according to the characteristics of the traffic flow in the building, and to which of the traffic patterns the current elevator traffic condition corresponds. Determining means, and switching means for selecting only one of the arithmetic means corresponding to the determination result of the determining means, wherein the learning data creating means determines the determination result of the determining means. The learning means creates learning data for each traffic pattern based on Corresponding weighting factor calculation means modifies each elevator control device and controls the car on the basis of the output data of the calculating means selected by said switching means.